Challenges in Building Large-Scale Information Retrieval...

Challenges in Building Large-Scale Information Retrieval Systems

Jeff Dean Google Senior Fellow jeff@google.com

• Challenging blend of science and engineering –Many interesting, unsolved problems –Spans many areas of CS:

• architecture, distributed systems, algorithms, compression, information retrieval, machine learning, UI, etc.

–Scale far larger than most other systems !

• Small teams can create systems used by hundreds of millions

Why Work on Retrieval Systems?

• Must balance engineering tradeoffs between: –number of documents indexed –queries / sec – index freshness/update rate –query latency – information kept about each document –complexity/cost of scoring/retrieval algorithms !

• Engineering difficulty roughly equal to the product of these parameters !

• All of these affect overall performance, and performance per $

Retrieval System Dimensions

• # docs: ~70M to many billion • queries processed/day: • per doc info in index: • update latency: months to minutes • avg. query latency: <1s to <0.15s !

• More machines * faster machines:

1999 vs. 2014

~100X

~3X~1000X

~10000X~6X

~1000X

• Parameters change over time –often by many orders of magnitude !

• Right design at X may be very wrong at 10X or 100X – ... design for ~10X growth, but plan to rewrite before ~100X !

• Continuous evolution: –7 significant revisions in last 10 years –often rolled out without users realizing we’ve made

major changes

Constant Change

• Evolution of Google’s search systems –several gens of crawling/indexing/serving systems –brief description of supporting infrastructure

!

–Joint work with many, many people !

• Interesting directions and challenges

Rest of Talk

“Google” Circa 1997 (google.stanford.edu)

Research Project, circa 1997

Frontend Web Server

I0 I1 I2 IN

Index shards

D0 D1 DM

query

Index servers Doc servers

… …Doc shards

• By doc: each shard has index for subset of docs – pro: each shard can process queries independently – pro: easy to keep additional per-doc information – pro: network traffic (requests/responses) small – con: query has to be processed by each shard – con: O(K*N) disk seeks for K word query on N shards

Ways of Index Partitioning

• By word: shard has subset of words for all docs – pro: K word query => handled by at most K shards – pro: O(K) disk seeks for K word query – con: much higher network bandwidth needed

• data about each word for each matching doc must be collected in one place

– con: harder to have per-doc information

In our computing environment, by doc makes more sense

• Documents assigned small integer ids (docids) –good if smaller for higher quality/more important docs

• Index Servers: –given (query) return sorted list of (score, docid, ...) –partitioned (“sharded”) by docid – index shards are replicated for capacity –cost is O(# queries * # docs in index)

!

• Doc Servers –given (docid, query) generate (title, snippet) –map from docid to full text of docs on disk –also partitioned by docid –cost is O(# queries)

Basic Principles

“Corkboards” (1999)

Serving System, circa 1999

Frontend Web Server

I0 I1 I2 IN

I0 I1 I2 IN

I0 I1 I2 IN

Replicas …

…

Index shards

D0 D1 DM

D0 D1 DM

D0 D1 DMReplicas …

…Doc shards

query

Index servers

Cache servers

C0 C1 CK…

Doc servers

Ad System

• Cache servers: –cache both index results and doc snippets –hit rates typically 30-60%

• depends on frequency of index updates, mix of query traffic, level of personalization, etc

!

• Main benefits: – performance! 10s of machines do work of 100s or 1000s – reduce query latency on hits

• queries that hit in cache tend to be both popular and expensive (common words, lots of documents to score, etc.)

!

• Beware: big latency spike/capacity drop when index updated or cache flushed

Caching

• Simple batch crawling system –start with a few URLs –crawl pages –extract links, add to queue –stop when you have enough pages !

• Concerns: –don’t hit any site too hard –prioritizing among uncrawled pages

• one way: continuously compute PageRank on changing graph

–maintaining uncrawled URL queue efficiently • one way: keep in a partitioned set of servers

–dealing with machine failures

Crawling (circa 1998-1999)

• Simple batch indexing system –Based on simple unix tools –No real checkpointing, so machine failures painful –No checksumming of raw data, so hardware bit errors

caused problems • Exacerbated by early machines having no ECC, no parity • Sort 1 TB of data without parity: ends up "mostly sorted" • Sort it again: "mostly sorted" another way !

• “Programming with adversarial memory” –Led us to develop a file abstraction that stored

checksums of small records and could skip and resynchronize after corrupted records

Indexing (circa 1998-1999)

• 1998-1999: Index updates (~once per month): –Wait until traffic is low –Take some replicas offline –Copy new index to these replicas –Start new frontends pointing at updated index and serve

some traffic from there

Index Updates (circa 1998-1999)

• Index server disk: –outer part of disk gives higher disk bandwidth

Index Updates (circa 1998-1999)

Oct99 index

• Index server disk: –outer part of disk gives higher disk bandwidth

Index Updates (circa 1998-1999)

Oct99 index

Nov99 index

1. Copy new index to inner half of disk (while still serving old index)

2. Restart to use new index

• Index server disk: –outer part of disk gives higher disk bandwidth

Index Updates (circa 1998-1999)

Nov99 index

1. Copy new index to inner half of disk (while still serving old index)

2. Restart to use new index

3. Wipe old index

• Index server disk: –outer part of disk gives higher disk bandwidth

Index Updates (circa 1998-1999)

Nov99 index

Nov99 index

1. Copy new index to inner half of disk (while still serving old index)

2. Restart to use new index

3. Wipe old index

4. Re-copy new index to faster half of disk

• Index server disk: –outer part of disk gives higher disk bandwidth

Index Updates (circa 1998-1999)

Nov99 index1. Copy new index to inner half of disk (while still serving old index)

2. Restart to use new index

3. Wipe old index

5. Wipe first copy of new index

4. Re-copy new index to faster half of disk

• Index server disk: –outer part of disk gives higher disk bandwidth

Index Updates (circa 1998-1999)

Nov99 index1. Copy new index to inner half of disk (while still serving old index)

2. Restart to use new index

3. Wipe old index

5. Wipe first copy of new index

6. Inner half now free for building various performance improving data structures

4. Re-copy new index to faster half of disk

Pair cache: pre-intersected pairs of posting lists for commonly co-occurring query terms (e.g. “new” and “york”, or “barcelona” and “restaurants”)

Nov99 pair cache

Google Data Center (2000)

Google (new data center 2001)

Google Data Center (3 days later)

• Huge increases in index size in ’99, ’00, ’01, ... – From ~50M pages to more than 1000M pages

!

• At same time as huge traffic increases – ~20% growth per month in 1999, 2000, ... – ... plus major new partners (e.g. Yahoo in July 2000 doubled traffic

overnight) !

• Performance of index servers was paramount – Deploying more machines continuously, but... – Needed ~10-30% software-based improvement every month

Increasing Index Size and Query Capacity

Dealing with Growth

Frontend Web Server

query

Index servers

Cache servers

Ad System

Doc Servers

Cache ServersReplicas

Index shards

I0 I1 I2I0 I1 I2

I3I3

I0 I1 I2 I3

I4I4I4

I10…I10…I10…

I0 I1 I2 I3 I4 I10…

… …

I60…I60…I60…I60…

…

• Shard response time influenced by: –# of disk seeks that must be done –amount of data to be read from disk !

• Big performance improvements possible with: –better disk scheduling – improved index encoding

Implications

• Original encoding (’97) was very simple: !

–hit: position plus attributes (font size, title, etc.) !

–Eventually added skip tables for large posting lists !

• Simple, byte aligned format –cheap to decode, but not very compact –... required lots of disk bandwidth

Index Encoding circa 1997-1999

docid+nhits:32b hit: 16b hit: 16b … hit: 16bdocid+nhits:32bWORD

• Bit-level encodings: –Unary: N ‘1’s followed by a ‘0’ –Gamma: log2(N) in unary, then floor(log2(N)) bits –RiceK: floor(N / 2K) in unary, then N mod 2K in K bits

• special case of Golomb codes where base is power of 2

–Huffman-Int: like Gamma, except log2(N) is Huffman coded instead of encoded w/ Unary !

• Byte-aligned encodings: –varint: 7 bits per byte with a continuation bit

• 0-127: 1 byte, 128-4095: 2 bytes, ...

– ...

Encoding Techniques

Block-Based Index Format

• Block-based, variable-len format reduced both space and CPU

delta to last docid in block: varintencoding type: Gamma # docs in block: GammaN -1 docid deltas: Ricek codedN values of # hits per doc: GammaH hit attributes: run length Huffman encodedH hit positions: Huffman-Int encoded

block length: varint

• Reduced index size by ~30%, plus much faster to decode

Block 0 Block 1 …Block 2 Block NSkip table(if large)WORD

Block format (with N documents and H hits):

Byte aligned header

• Must add shards to keep response time low as index size increases !

• ... but query cost increases with # of shards –typically >= 1 disk seek / shard / query term –even for very rare terms !

• As # of replicas increases, total amount of memory available increases –Eventually, have enough memory to hold an entire

copy of the index in memory • radically changes many design parameters

Implications of Ever-Wider Sharding

Early 2001: In-Memory Index

Frontend Web Server

query

Index servers

Cache servers

Ad System

Doc Servers

Cache Servers

Index shards

Shard 0

I0 I1 I2

I14

I3

I12

bal

I4 I5…I13

Shard 1

I0 I1 I2

I14

I3

I12

bal

I4 I5…I13

Shard 2

I0 I1 I2

I14

I3

I12

bal

I4 I5…I13

Shard N

I0 I1 I2

I14

I3

I12

bal

I4 I5…I13

Balancers

…

• Many positives: –big increase in throughput –big decrease in latency

• especially at the tail: expensive queries that previously needed GBs of disk I/O became much faster

e.g. [ “circle of life” ]!!

• Some issues: –Variance: touch 1000s of machines, not dozens

• e.g. randomized cron jobs caused us trouble for a while

–Availability: 1 or few replicas of each doc’s index data • Queries of death that kill all the backends at once: very bad • Availability of index data when machine failed (esp for important docs):

replicate important docs

In-Memory Indexing Systems

• Many datacenters around the world

Google’s Computational Environment Today

Zooming In...

Zooming In...

Lots of machines...

Save a bit of power: turn out the lights...

Cool...

Serving Design, 2004 edition

Root

…

…

Parent Servers

……

Leaf Servers

Repository Shards

…

RepositoryManager

File Loaders

Cache servers

Requests

GFS

New Index Format

• Block index format used two-level scheme: – Each hit was encoded as (docid, word position in doc) pair – Docid deltas encoded with Rice encoding – Very good compression (originally designed for disk-based indices), but

slow/CPU-intensive to decode !

• New format: single flat position space – Data structures on side keep track of doc boundaries – Posting lists are just lists of delta-encoded positions – Need to be compact (can’t afford 32 bit value per occurrence) – … but need to be very fast to decode

Byte-Aligned Variable-length Encodings

• Varint encoding: – 7 bits per byte with continuation bit – Con: Decoding requires lots of branches/shifts/masks

0000111100000001 0000001111111111 1111111111111111 00000111

1 15 511 131071

• Idea: Encode byte length as low 2 bits – Better: fewer branches, shifts, and masks – Con: Limited to 30-bit values, still some shifting to

decode

0000111100000001 0000001101111111 1111111110111111 00000111

1 15 511 131071

Group Varint Encoding• Idea: encode groups of 4 values in 5-17 bytes

– Pull out 4 2-bit binary lengths into single byte prefix

0000111100000001 0000001101111111 1111111110111111 00000111

0000111100000001 11111111 11111111

1 15 511 131071

00000001 111111110000000100000110

Tags

• Much faster than alternatives: – 7-bit-per-byte varint: decode ~180M numbers/second – 30-bit Varint w/ 2-bit length: decode ~240M numbers/second – Group varint: decode ~400M numbers/second

• Decode: Load prefix byte and use value to lookup in 256-entry table:

00000110 Offsets: +1,+2,+3,+5; Masks: ff, ff, ffff, ffffff……

2007: Universal Search

Frontend Web Server

query

Cache servers

Ad System

News

Super root

Images

WebBlogs

VideoBooks

Local

Indexing Service

• Low-latency crawling and indexing is tough –crawl heuristics: what pages should be crawled? –crawling system: need to crawl pages quickly – indexing system: depends on global data

• PageRank, anchor text of pages that point to the page, etc. • must have online approximations for these global properties

–serving system: must be prepared to accept updates while serving requests • very different data structures than batch update serving system

Index that? Just a minute!

Often use hybrid systems (1+2): 1. Very fast update system, but with modest capacity (high $$$/doc/query) 2. Slower updating system, but with huge capacity and relatively immutable data structures (low $/doc/query)

Understanding Text

!… car parking available for a small fee. … parts of our floor model inventory for sale.

Document 1

![ car parts for sale ]

Query

!Selling all kinds of automobile and pickup truck parts, engines, and transmissions.

Document 2

• Embeddings & Neural Nets

• How we use them

• What’s next?

go/brain

dolphin

Sea World

Paris

N-dimensional Embeddings

porpoise

Camera

Embedding Function: A look-up-table that maps sparse features into dense floating point vectors.

3-D embedding space

Usually use many more dimensions than 3 (e.g. 1000-D embeddings)

ESingle embedding function

Hierarchical softmax classifier

Raw sparse features Obama’s

nearby word

Skipgram Text Model

meeting with Putin

Mikolov, Chen, Corrado and Dean. Efficient Estimation of Word Representations in Vector Space, http://arxiv.org/abs/1301.3781

Open source implementation: https://code.google.com/p/word2vec/

source word

nearby words

embedding! vector

upper layers

Nearest neighbors in language embeddings space are closely related semantically.

tiger shark!!bull shark!blacktip shark!shark!oceanic whitetip shark!sandbar shark!dusky shark!blue shark!requiem shark!great white shark!lemon shark

car!!cars!muscle car!sports car!compact car!autocar!automobile!pickup truck!racing car!passenger car !dealership

new york!!new york city!brooklyn!long island!syracuse!manhattan!washington!bronx!yonkers!poughkeepsie!new york state

• Trained language model on Wikipedia corpus.

E

* 5.7M docs, 5.4B terms, 155K unique terms, 500-D embeddings

Embeddings are Powerful

man

woman

king

queen

swam

walking

swimming

walked

Solving Analogies

• Embedding vectors trained for the language modeling task have very interesting properties (especially the skip-gram model).

!

E(hotter) - E(hot) ≈ E(bigger) - E(big) !

E(Rome) - E(Italy) ≈ E(Berlin) - E(Germany)

!

E(hotter) - E(hot) + E(big) ≈ E(bigger) !

E(Rome) - E(Italy) + E(Germany) ≈ E(Berlin)

Details in: Efficient Estimation of Word Representations in Vector Space. Mikolov, Chen, Corrado and Dean. NIPS 2013.

Solving Analogies

• Embedding vectors trained for the language modeling task have very interesting properties (especially the skip-gram model).

Skip-gram model w/ 640 dimensions trained on 6B words of news text achieves 57% accuracy for analogy-solving test set.

Embeddings are Powerful

Spain

Italy

Germany

Turkey

Russia

Canada

Japan

Vietnam

China BeijingHanoi

Tokyo

OttawaMoscow

Ankara

Berlin

Rome

Madrid

fallen

draw

fell

drawn

drew taketaken

took

givegiven

gave

fall

N-dimensional Embeddings

Plenty of applications:

• Machine Translation

• Synonym handling

• Sentiment analysis

• Ads

• Semantic visual models

• Entity disambiguation

Embedding Longer Pieces of Text

Query Doc

Embedding Longer Pieces of Text

Englishsentence

Cantonese sentence

Swahili sentence

Embedding Longer Pieces of Text

word word word word word

sentence

word word word word word

RNN / LSTM TreeNN

sentence

EEmbedding function

Deep neural network

Raw sparse inputs

Floating-point vectors

Prediction (classification or regression)

Using Embeddings in Larger Neural Models

features

E1 E2 E3 E4Separate embedding

functions

Raw feature categories f1 f2 f3 f4

words!in!

creative

user!country

words!in!

query

...!

Logistic regressionpCTR

Predicting clicks

click?

Embeddings and Neural Nets Show Considerable Promise

In Conclusion...

• Designing and building large-scale retrieval systems is a challenging, fun endeavor –new problems require continuous evolution –work benefits many users –new retrieval techniques often require new systems !

!

• Thanks for your attention!

Thanks! Questions...?

• Further reading: Ghemawat, Gobioff, & Leung. Google File System, SOSP 2003.

Barroso, Dean, & Hölzle. Web Search for a Planet: The Google Cluster Architecture, IEEE Micro, 2003.

Dean & Ghemawat. MapReduce: Simplified Data Processing on Large Clusters, OSDI 2004.

Chang, Dean, Ghemawat, Hsieh, Wallach, Burrows, Chandra, Fikes, & Gruber. Bigtable: A Distributed Storage System for Structured Data, OSDI 2006.

Brants, Popat, Xu, Och, & Dean. Large Language Models in Machine Translation, EMNLP 2007.

Mikolov, Chen, Corrado and Dean. Efficient Estimation of Word Representations in Vector Space, http://arxiv.org/abs/1301.3781 Dean, Corrado, et al. , Large Scale Distributed Deep Networks, NIPS 2012. !

• These and many more available at: http://labs.google.com/papers.html http://research.google.com/people/jeff